evidence: GPU beta scorecard — Kakeya vs standalone AR on H200 (main #117)#119
Merged
Merged
Conversation
…emory/context/throughput k3_e2e_gpu_bench (NIAH 160/320 lines, incremental restored vs AR): - memory: AR KV 733/1454 MB vs Kakeya constant 16.71 MB -> 43.9x/87.0x saving - context: 68-tok window reconstructs 3254/6454 tok -> 47.9x/94.9x, recall 1.0==AR - decode: restored 16.5 vs AR 16.3-16.6 tok/s (~1.01x parity) k3_specdecode_gpu_bench (block-16, 3238-tok prompt): - AR 16.125 | restored-pertoken 16.297 (1.01x) | FUSED 28.937 tok/s (1.79x AR) - recall 1.0 across AR/pertoken/fused (byte-identical), accept_len 3.32 Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
cursor Bot
pushed a commit
that referenced
this pull request
Jun 14, 2026
…pology test record Case 1 (RUN): gRPC RuntimeService on Mac M4 sustains 256/256 concurrent agent connections, zero errors; per-session KV bounded 7.80 MB; node KV upper bound ~2.0 GB; server RSS flat ~3.85 GB. Single-tenant caveat: generate serializes (latency linear in N) -> 256 = max concurrent connections served, not parallel inferences. Evidence: results/research/k3_agent_capacity_mac.json. Case 2 (FEASIBILITY): cross-host GPU-proposer<->Mac-verifier discovery+draft is design-only (no distributed.proto / CapabilityService / ProposeBlock) AND WAN-bounded out by latency. Realizable topology: WAN=control/tool plane (bridge), LAN=co-located data plane. Proxies: GPU 1.79x AR (#119), Mac 0.93x (#118), max conns 256+ (Case 1), bounded Mac KV. Also records served-MLX-gemma gap. Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
cursor Bot
pushed a commit
that referenced
this pull request
Jun 14, 2026
…easibility (Case 2) — ADR 0014 (#123) * test(case1): gRPC agent-connection capacity load test + Mac-bridge preset scripts/research/grpc_agent_capacity_loadtest.py launches a RuntimeService subprocess and ramps N concurrent agents (independent gRPC channel + session each), reporting max concurrent agents, per-session bounded KV (GetSessionInfo), node KV upper bound (capacity * per-session bound), create/generate latency curve, and server RSS. Honest about v0.3 single-tenant (shared verifier, RPCs serialized) — measures connection/admission scaling, not parallel inference. New manifest preset 'agent-capacity-loadtest' runs it on the Mac's real MLX gemma verifier. Validated locally on the cloud agent (cpu Qwen3-1.7B). Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com> * test(case1): use cpu Qwen3-0.6B for capacity preset (served MLX gemma path is a v0.4 gap; connection scaling is model-independent) Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com> * test(case2)+adr: ADR 0014 — agent-connection capacity & cross-host topology test record Case 1 (RUN): gRPC RuntimeService on Mac M4 sustains 256/256 concurrent agent connections, zero errors; per-session KV bounded 7.80 MB; node KV upper bound ~2.0 GB; server RSS flat ~3.85 GB. Single-tenant caveat: generate serializes (latency linear in N) -> 256 = max concurrent connections served, not parallel inferences. Evidence: results/research/k3_agent_capacity_mac.json. Case 2 (FEASIBILITY): cross-host GPU-proposer<->Mac-verifier discovery+draft is design-only (no distributed.proto / CapabilityService / ProposeBlock) AND WAN-bounded out by latency. Realizable topology: WAN=control/tool plane (bridge), LAN=co-located data plane. Proxies: GPU 1.79x AR (#119), Mac 0.93x (#118), max conns 256+ (Case 1), bounded Mac KV. Also records served-MLX-gemma gap. Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com> * test(case1): harness --context-len + FD-limit raise; add agent-capacity-stress preset (ramp to 2048, probe connection + memory ceiling) Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com> * test(case1): finer stress levels to pinpoint single-tenant serialization knee; save cap-2048 stress evidence Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com> * test(case2): cross-host RTT-sweep mode — inject per-block proposer<->verifier round-trip into the fused engine to measure the WAN-penalty throughput curve Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com> * docs: sync ADR 0014 (agent-connection capacity + cross-host topology) into README Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com> * test: Case-1 stress ceilings + Case-2 measured cross-host WAN-penalty curve Case 1 stress (Mac M4): FD not the limit (100k); memory scales with capacity x window (cap 2048 -> 11.5GB, node bound 61GB > 24GB RAM); single-tenant serialization caps heavy-context concurrency at ~8 (vs 256 light). Case 2 (H200 real models): injected per-block proposer<->verifier RTT -> measured WAN-penalty curve. Co-located 2.20x AR; break-even ~100ms/block; 150ms -> 0.77x (net loss). LAN (<=15ms) keeps 1.8-2.2x. Confirms WAN data plane infeasible. Updates ADR 0014 + README with measured curves + evidence JSONs. Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com> --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Pulls the GPU beta from
main(9d5e6b4, #107 fused engine + #117 consolidation) onto an H200 (vast.ai) and runs the Kakeya inference engine vs standalone Gemma-4 26B AR scorecard — the GPU analog of the Mac "mlx-only" comparison. Commits the two evidence JSONs.Verifier
google/gemma-4-26B-A4B-it(bf16), drafterz-lab/gemma-4-26B-A4B-it-DFlash,f_theta_v5_s5_sliding, S5 (5 exact full-attn layers).Scorecard (NVIDIA H200,
main@9d5e6b4)1) Memory bounded — resident KV
Kakeya KV is constant 16.71 MB (68-tok sink+window) regardless of context; AR grows linearly → saving scales with context.
2) Context length — bounded window vs effective context
3) Token throughput (decode tok/s, 3238-tok prompt)
block-16, accept_len 3.32, byte-identical output.
Net (GPU): bounded memory (44–87× KV saving, constant 16.71 MB) + full-context recall (48–95× compression) + 1.79× AR throughput, all at AR-identical correctness. Confirms the platform fork: spec-decode value lands on GPU (cheap verify-batch) vs Mac ~0.93× (26B
verify(L)dominates).Files
results/research/k3_gpu_beta_e2e_memory_context.json—k3_e2e_gpu_bench(memory/context/recall, rungs 160/320).results/research/k3_gpu_beta_fused_throughput.json—k3_specdecode_gpu_bench(AR / restored / fused throughput).Testing
k3_e2e_gpu_bench.py --haystack-lines 160,320 --incremental(H200, exit 0, recall 1.0 both rungs)k3_specdecode_gpu_bench.py --block-size 16 --skip-unfused(H200, exit 0, fused 1.79× AR, recall 1.0)